37 tok/s from Qwen3-Next-80B (48 GB) on a 32 GB laptop, 24 without drafts vs 2.8 for llama.cpp; reading 62 vs 16 tok/s. Faster than llama.cpp on all 13 models compared in Q4_K_M. LLMs bigger than your RAM, pure C, CPU or GPU.
C avx2ccpu-inferenceggufinference-enginelinuxllmlocal-llmmixture-of-expertsmoenvmequantizationqwen3simdspeculative-decodingvulkan
A chat in FreeBASIC on Janas-LLM
The smallest chat on Janas-LLM, written twice:
chat.bas, procedure by procedure, withinclude/janas/llm.bi: the C library's functions as they are, for FreeBASIC and for BASIC MODERN, the dialect Prabanta implements.chat_oop.bas, with the classes ofinclude/janas/janas_oop.bi: aJanas.Model, itsJanas.Chat, and a handler that extendsJanas.Printerto print the reply as it is written and its counters at the end. Nothing is closed by hand: the objects close what they hold when the program ends.
Both read the user's lines from a terminal, or from a pipe.
Building and running
Janas built first (./build.sh in the repository's root), which makes the shared libraries in bin/x86_64-linux/; then, from the root, with FreeBASIC 1.10:
fbc examples/fb-chat/chat_oop.bas -i include -p bin/x86_64-linux -x janas-chat-oop
LD_LIBRARY_PATH=bin/x86_64-linux ./janas-chat-oop models/jns/qwen3-4b-q4km.jns
echo "Ciao" | LD_LIBRARY_PATH=bin/x86_64-linux ./janas-chat-oop models/jns/qwen3-4b-q4km.jns
and the same for chat.bas. The classes are described in docs/library.md, "From FreeBASIC and BASIC MODERN".